thatjoeoverthr

Curiously, Star Trek I think had Data intended as an artificial person, in a context where AGI is already normal. The computers are depicted with significant AI capabilities including analysis, question-answering, generation, chat interfaces, and the holodeck (their favourite toy) is substantially better than Data at human imitation. Nobody seems to be confused about it, or especially impressed. One of the holodeck episodes centres on the holodeck outwitting Data specifically, after they inadvertently prompt it to do so. Part of Data's deal is he actually has to work his way up as a fully embodied, physically limited artificial man with personal ambitions. Really interesting to view this in hindsight from 2026!

Random things added constantly. TBH a bit distressing.

The 5-6 months is in line with other slop brands I’ve tried, e.g. H&M. Sad to see Levis join them in the gutter.

While there are ways to manage it with supervision, the impulses of every model are clear and counterproductive. All code is littered CoT in comments and insane decisions. Its attention is limited and it saturates fast, neutralizing one of its main advantages. It’s really nice with supervision but something smells with the hustlers.

Look, you're must not saying anything. "If done right ... you probably wouldn’t care". Yes, correct. It's done bad, so I don't like it, and the visible slack is a sign of slack. Would you like to push back in some way for some reason? What's the call to action here?

Nice, now add slackers and non-writers to the equation.

Quick tip to make AI code better is just delete the comments. Usually they’re CoT that shouldn’t be there.

If

Stretching nothing, but encompassing a fairly standard payload. “The crisis of men is they don’t trust us, the authorities.” No need to pin down why. We’ve poisoned the well so the article is done.

But I do know them, and so do you, and they’re signs of cost cutting partly for the reason you mention. They’re the careless, basic “no-brainer” choice of models, plus a lack of editing, or a very light “edit” in which obvious, even famous tics are left in. (Many real editors would clean up writers’ tics.) So it’s possible care was put into it, but tics visibility is evidence against it and not evidence for it. If your text does pass and educated readers don’t see it, it’s a consequence of care and intent, which is what we’re really asking for.

This is logical, but it’s also a lemon market. You don’t know the author, editor and publisher did anything of the sort, unless you’re the sort of person who doesn’t need the book. But if you do see an LLMism, it’s a strong sign of cost-cutting.

This, and the fact that GPT is writing for you personally. A text written to persuade a publisher, or to manipulate Google results, has the deck stacked hard against it vs. asking even a cheap AI that can size you up personally.

This is the bar with other works as well. If you would hire it buy anything that’s just a mindless AI dump, why not use your $5/mo GPT sub?

This goes back far, and I’ve even seen what appear to be shills for random B-list celebrities. Add SEO business as usual, political campaigns, international influence and the SnR is just radio static.

No, I’m referring to something else. Not the next word task. Being able to know whether you know something. Next time you encounter a child, ask them something they’re likely to be ignorant of, and see they’re able to say “I don’t know”.

“trip themselves in their attempts at achieving a task”

I see this a lot in Claude Code. I assume it has to do with the training structure.

Example is “fallbacks”. Claude constantly sprinkles “fallbacks” in the code, even when I ask it not to. That is, write multiple candidate implementations into the same code with some kind of switch.

This is a problem because you only need one, and it would seem to have inflated the code for no reason. (The madness accelerates with code volume, so you must push back.) anyway, few people would do it this way.

But I thought, what could be the benefit?

If you’re being conditioned to pass evals one-shot with code that will be discarded and never read, it’s a great strategy. If you have more than one way to solve it, you can just put both. The behavior would easily be reinforced, if trained that way.

But in any case, again, a certain nature and certain conditioning.

I think we’ll learn to accept it as AGI but also that no intelligence is fully divorced from context, limits and conditioning.

It’s really not like that. If I ask you to tell me about a geographical place I just made up, you can trivially and generally instantly recognize you don’t recognize it. A child can do this. You wouldn’t be able to hold a job or generally get through life without this level of self awareness.

Their lack of self reference is a core problem that undergirds a lot of faults that do occur during inference, but their breadth + the agent harness successfully covers it well, so it requires a bit of poking to witness. The “hallucination” phenomenon is exactly this. They don’t know the scope of their own knowledge, and they just say stuff, so if you go out of band, it has a higher probability emitting claims that aren’t true. RAG (I don’t mean embedding indices, but any information ingest such as an agent harness executing a search) are somewhat effective in covering for it, enough to make them very useful! But when it does go wrong, it’s generally the same reasons. It has a certain nature and sometimes you run afoul of it.

But I suppose it doesn’t harm its reasoning!

Alas, he is not, and two things are true for him:

1. AI models produce consistent output, so it looks the same everywhere, which makes you look like a hustler if you use it for copywriting or design.

2. Customers who are fine with this can use Lovable for $20/month, so it's loser's game to chase those customers.

> always picks its own

If you hate AI writing enough, this turns AI filters into a kind of humiliation ritual. AI will derank normal business writing for human readers, and uprank inflated, verbose, tic-heavy slop. So you have to put the heavy slop out with your name on it. Really perverse moment.

“let the embedding be a latent space that forms naturally when training the network”

The embeddings are produced in concert with the network, to serve the network, and not created as a separate step.

It’s actually very cool

The look-up table is a matrix. Each row is an embedding and each row number is a token ID.

You get a differentiable transformation from token ID to token embedding using a “one hot vector” and a matrix multiplication

If you take the transpose of this matrix, you can convert an internal representation back to the same token form, but treat it as logits and give it to the sampler.

So token embeddings are produced on demand in service of the model, according to the model’s needs.

I found an example of this strategy in a paper as far back as 1980!

In the other reply I recommend the Bengio paper. But do bite the bullet and try it.

Embeddings are there to make continuous-space identities so you can run them through a differentiable model. Without this the tokens (no matter your granularity) are pure surrogate identities, and you can’t run a gradient through them. You also hit the curse of dimensionality hard because the model can’t perceive similarity. “Cat” and “kitten” for example are simply different atoms of text, but with embeddings, you can leverage what you learned about “cat” when you encounter “kitten”. Look at “A Neural Probabalistic Language Model” (Bengio, 2003).

You can actually rig up an embedding variant of a Markov chain with just a few tokens of context, and no position coding, transformers, attention, none of it, and only minutes of training time. As long as you have the embedding lookup table trainable it will do some neat stuff.

I migrated a client off for the same reason. The “app platform” ran behind Cloudflare, but they run the Cloudflare, so I couldn’t put it behind my own Cloudflare, nor could I configure it and solve the problem. Moved the entire thing to Hetzner, fast. What a week.

Nobody thinks customer service agents are real, realistic or useful. EU just passes laws.

Thanks!

Few reputations remain unruined now.

You have to be jacked to use the jackhammer.

Yes, this. Comprehension is the point. We could map this to something like physics. If a man on a horse can shoot another man with a bow, empirically he makes correct predictions on gravity, wind and relative motion. But he can’t explain it. It’s not any different if your model has some “embodied” or demonstrable understanding; the model is not part of the discourse.

One problem here with the protocols is the field isn't yet so stable. Most people are copying OpenAI's endpoint, and so you have "OpenAI compatible", but there are always little hairs on it. Case in point, OpenAI itself no longer offers multiple basic inference options on their new models. Can't pass temp, get back logits, etc. on anything after GPT 4.1. It's not a total zoo and Open Router is pretty easy but everything is just enough a special snowflake that models aren't a commodity. Maybe later though.